Update README.md
Browse files
README.md
CHANGED
|
@@ -9,15 +9,65 @@ model-index:
|
|
| 9 |
- name: LLaMA-3-8B-Instruct-TR-DPO
|
| 10 |
results:
|
| 11 |
- task:
|
| 12 |
-
type:
|
| 13 |
dataset:
|
| 14 |
-
type:
|
| 15 |
name: MMLU_TR_V0.2
|
| 16 |
metrics:
|
| 17 |
- name: 5-shot
|
| 18 |
type: 5-shot
|
| 19 |
value: 0.4983
|
| 20 |
verified: false
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
---
|
| 22 |
|
| 23 |
<img src="https://huggingface.co/Metin/LLaMA-3-8B-Instruct-TR-DPO/resolve/main/llama.png"
|
|
@@ -101,12 +151,14 @@ print(outputs[0]["generated_text"][len(prompt):])
|
|
| 101 |
|
| 102 |
## OpenLLMTurkishLeaderboard_v0.2 benchmark results
|
| 103 |
|
| 104 |
-
MMLU_TR_V0.2
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
|
|
|
|
|
|
| 110 |
|
| 111 |
## Output Example (DPO Model vs Base Model)
|
| 112 |
|
|
|
|
| 9 |
- name: LLaMA-3-8B-Instruct-TR-DPO
|
| 10 |
results:
|
| 11 |
- task:
|
| 12 |
+
type: multiple-choice
|
| 13 |
dataset:
|
| 14 |
+
type: multiple-choice
|
| 15 |
name: MMLU_TR_V0.2
|
| 16 |
metrics:
|
| 17 |
- name: 5-shot
|
| 18 |
type: 5-shot
|
| 19 |
value: 0.4983
|
| 20 |
verified: false
|
| 21 |
+
- task:
|
| 22 |
+
type: multiple-choice
|
| 23 |
+
dataset:
|
| 24 |
+
type: multiple-choice
|
| 25 |
+
name: Truthful_QA_V0.2
|
| 26 |
+
metrics:
|
| 27 |
+
- name: 0-shot
|
| 28 |
+
type: 0-shot
|
| 29 |
+
value: 0.5232
|
| 30 |
+
verified: false
|
| 31 |
+
- task:
|
| 32 |
+
type: multiple-choice
|
| 33 |
+
dataset:
|
| 34 |
+
type: multiple-choice
|
| 35 |
+
name: ARC_TR_V0.2
|
| 36 |
+
metrics:
|
| 37 |
+
- name: 25-shot
|
| 38 |
+
type: 25-shot
|
| 39 |
+
value: 0.4437
|
| 40 |
+
verified: false
|
| 41 |
+
- task:
|
| 42 |
+
type: multiple-choice
|
| 43 |
+
dataset:
|
| 44 |
+
type: multiple-choice
|
| 45 |
+
name: HellaSwag_TR_V0.2
|
| 46 |
+
metrics:
|
| 47 |
+
- name: 10-shot
|
| 48 |
+
type: 10-shot
|
| 49 |
+
value: 0.4558
|
| 50 |
+
verified: false
|
| 51 |
+
- task:
|
| 52 |
+
type: multiple-choice
|
| 53 |
+
dataset:
|
| 54 |
+
type: multiple-choice
|
| 55 |
+
name: GSM8K_TR_V0.2
|
| 56 |
+
metrics:
|
| 57 |
+
- name: 5-shot
|
| 58 |
+
type: 5-shot
|
| 59 |
+
value: 0.5421
|
| 60 |
+
verified: false
|
| 61 |
+
- task:
|
| 62 |
+
type: multiple-choice
|
| 63 |
+
dataset:
|
| 64 |
+
type: multiple-choice
|
| 65 |
+
name: Winogrande_TR_V0.2
|
| 66 |
+
metrics:
|
| 67 |
+
- name: 5-shot
|
| 68 |
+
type: 5-shot
|
| 69 |
+
value: 0.5506
|
| 70 |
+
verified: false
|
| 71 |
---
|
| 72 |
|
| 73 |
<img src="https://huggingface.co/Metin/LLaMA-3-8B-Instruct-TR-DPO/resolve/main/llama.png"
|
|
|
|
| 151 |
|
| 152 |
## OpenLLMTurkishLeaderboard_v0.2 benchmark results
|
| 153 |
|
| 154 |
+
- **MMLU_TR_V0.2**: 49.83%
|
| 155 |
+
- **Truthful_QA_TR_V0.2**: 52.32%
|
| 156 |
+
- **ARC_TR_V0.2**: 44.37%
|
| 157 |
+
- **HellaSwag_TR_V0.2**: 45.58%
|
| 158 |
+
- **GSM8K_TR_V0.2**: 54.21%
|
| 159 |
+
- **Winogrande_TR_V0.2**: 55.06%
|
| 160 |
+
|
| 161 |
+
These scores may differ from what you will get when you run the same benchmarks, as I did not use any inference engine (vLLM, TensorRT-LLM, etc.)
|
| 162 |
|
| 163 |
## Output Example (DPO Model vs Base Model)
|
| 164 |
|